Reconstructing Regularly Missing Seismic Traces With a Classifier-Guided Diffusion Model
Xinlei Wang, Zhiguo Wang, Zhe Xiong, Yang Yang, Chaobo Zhu, Jinghuai Gao
Abstract
Reconstructing missing seismic data is crucial for seismic processing and interpretation. Recent methods struggle when seismic traces are regularly missing, such as near offset data. We proposed a classifier-guided conditional seismic denoising diffusion probabilistic model (CCSeis-DDPM) to enable consistent reconstructions. The CCSeis-DDPM adopts the Markov model architecture of denoising diffusion probabilistic models to generate high-quality results. The model involves classifier-guided training and tailored inference. During training, we employ a U-Net with embedded timestep and three class labels for noise prediction, utilizing classifier guidance to enhance reconstruction accuracy. In the inference phase, the model selectively samples unmasked regions using available seismic data. Our experiments on synthetic and field shot gathers with regularly missing near, mid and far offsets show the proposed CCSeis-DDPM reconstructs regularly missing traces more accurately than current state-of-the-art methods, demonstrated qualitatively and quantitatively. This successful integration of diffusion probabilistic models with classification guidance and conditioning underscores the immense potential of this approach for enhancing seismic data reconstruction processes.